@inproceedings{pan2023dorewards,
    title        = {{Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the \emph{Machiavelli} Benchmark}},
    author       = {Pan, Alexander and Shern, Chan Jun and Zou, Andy and Li, Nathaniel and Basart, Steven and Woodside, Thomas and Ng, Jonathan and Zhang, Hanlin and Emmons, Scott and Hendrycks, Dan},
    year         = {2023},
    booktitle    = {International Conference on Machine Learning},
    pages        = {26837--26867},
    organization = {PMLR}
}